Re: [BUG] sched/fair: divide error in __calc_prop_weight() from the enqueue path (flat-hierarchy series)

[Jake Steinman]

Another report of this, same divide but reached from the idle path rather
than fork. https://github.com/CachyOS/linux-cachyos/issues/985

  Oops: divide error: 0000 [#1] SMP NOPTI
  CPU: 7 UID: 0 PID: 0 Comm: swapper/7 Tainted: G         C          7.2.0-1-cachyos #1
  RIP: 0010:enqueue_task_fair+0x87c/0x2150
  Call Trace:
   <TASK>
   sched_ttwu_pending+0x204/0x590
   flush_smp_call_function_queue+0x169/0x250
   ? cpuidle_enter_state+0xc2/0x230
   cpu_startup_entry+0x2f6/0x3d0
   start_secondary+0x9a/0xa0
   common_startup_64+0x13e/0x150
   </TASK>
  Kernel panic - not syncing: Attempted to kill the idle task!

An idle CPU takes a wakeup IPI and faults enqueuing the woken task. In
swapper there is nothing to kill, so it panics immediately instead of
oopsing first.

Symbolised against linux-cachyos-dbg it lands where mine did:

  __calc_prop_weight   fair.c:4728:10
    inlined at enqueue_hierarchy  fair.c:7891:12
    inlined at enqueue_task_fair  fair.c:7949:11

Still nothing on what empties the cpuset. Their machine looks like mine —
concur mode, no empty effective masks, no partitions.

They have offered to run the tg_cpus() floor or a debug build with
WARN_ON_ONCE(!cfs_rq->load.weight). They reproduce more readily than I do,
so that is probably the machine to instrument.

Thanks,
Jake